Solutions · Reasoning
Best LLMs for deep reasoning
Frontier math, planning, and multi-step problem solving.
Reasoning routes weight GPQA Diamond, AIME, and MATH-500 heavily and permit longer thinking budgets. Use auto:reasoning when the answer matters more than latency — research, analysis, financial modeling, and complex planning.
| # | Model | Score | GPQA Diamond | Why |
|---|---|---|---|---|
| 1 | gpt-5.5 | 97 | 84.1% | Top-line reasoner; extended thinking crushes olympiad math. |
| 2 | claude-opus-4.5 | 96 | 82.7% | Best natural-language proofs and long-form analysis. |
| 3 | gemini-3.1-pro | 95 | 82.0% | Massive context + strong chain-of-thought on scientific PDFs. |
| 4 | o3 | 94 | 81.2% | Deliberate reasoner; slow but very high AIME pass rate. |
| 5 | grok-5 | 91 | 78.3% | Excellent on quantitative reasoning with tool use. |
| 6 | deepseek-r1 | 90 | 77.5% | Open-weights R1 lineage; strongest OSS math model. |
| 7 | qwen3-235b-thinking | 88 | 75.1% | Thinking-mode Qwen; great value on hard STEM. |
| 8 | magistral-medium-1.2 | 84 | 71.4% | Mistral's dedicated reasoner; solid AIME performance. |
| 9 | o4-mini | 83 | 70.8% | Cheap reasoning at scale for batched analytical work. |
| 10 | qwq-32b | 79 | 66.2% | Open, self-hostable reasoner for regulated workloads. |
Why these models
- Composite of GPQA Diamond, AIME 2025, MATH-500, and MMLU-Pro.
- Extended-thinking variants preferred where available.
- Cost is a soft signal — the router will trade $ for accuracy here.
Drop-in example
const res = await client.chat.completions.create({
model: "auto:reasoning",
reasoning: { effort: "high" },
messages: [{
role: "user",
content: "Prove that every prime > 3 is congruent to 1 or 5 mod 6.",
}],
});